Reproduce “EfficientVLM: Fast and Accurate Vision-Language Models via Knowledge Distillation and Modal-adaptive Pruning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tiannan Wang, Wangchunshu Zhou, Yan Zeng, Xinsong Zhang
EfficientVLM compresses the X-VLM vision-language model by distilling logits, attention, and hidden states into a smaller Transformer student, then learning task-specific pruning masks across vision, text, and cross-modal encoders. It retains 98.4% of the teacher’s average benchmark performance, with about 92–93 million parameters and measured speedups of 1.9× on GPU and 2.2× on CPU, but these are hardware-specific latency results, not universal efficiency gains or reduced training computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.